ic-metrics

Allocation-free, dependency-free measurement arithmetic.

The no_std library exports MeasurementAggregate for mergeable count/total/maximum with sticky actual-overflow flags, and MeasurementSummary for ordered observations with a latest value. MeasurementHistogram adds a mergeable fixed-size distribution with caller-selected bounds. Zero is a valid observation; an empty summary has no latest or maximum value. Tracked count and total clip independently, with sticky actual-overflow flags. Counts and individual samples use u64; totals and means use u128. Every exact counted batch fits its total. An exactly reached numeric cap remains exact. Consumers establish units, identity and reset boundaries before comparing observations; saturated totals cannot supply exact interval arithmetic.

Pending 0.8.2 adds MeasurementCounter for scalar event, item or byte totals. It holds a u128 total and sticky actual-overflow flag, with add, merge, validated from_parts, exact_total and identity-qualified delta_since. Zero is an exact scalar value; an absent metric stays consumer-owned. An exactly reached cap remains exact, and narrow reports use checked conversions. This counter retains no observation statistics and has no stable-store encoding.

use ic_metrics::MeasurementSummary;

let mut summary = MeasurementSummary::EMPTY;
assert_eq!(summary.latest(), None);
summary.record(0);
assert_eq!(summary.samples(), 1);
assert_eq!(summary.latest(), Some(0));

Tracked MeasurementAggregate::mean() and MeasurementSummary::mean() return Result<Option<u128>, MeasurementMeanError>: empty is None, measured zero is Some(0), and exact means round down. Actual overflow returns a typed error; exactly reaching the cap remains valid. scaled_mean(scale) checks the same flags and retains fractional information: samples 3 and 4 yield 35 at scale ten. Accumulator flags cannot restore a clamped input sample.

MeasurementAggregate::merge combines count/total and maximum without claiming ordering or deduplication. interval_since(previous) returns checked count/total differences and mean after the owner establishes common identity, units and append-only continuity. It rejects actual overflow and numeric discontinuity; maximum and latest values cannot be subtracted. Ordered summaries do not merge.

The published 0.8.0 hard cut removes record_sample, raw-field checked_mean and MeasurementMeanError::TotalWithoutSamples. Record into a tracked aggregate, summary or histogram; project consumer-owned reports from that state and preserve its overflow flags or explicit unavailable results. Means return a typed Projection error if the exact ratio owner rejects a scaled result. Units, identity, attribution, reset windows and any once-per-window alert remain local. The stable V1 bytes, map/key/epoch contract and allocation lifecycle are unchanged; this cut requires no stored-data reset. The dependency example selects the published 0.8 contract.

The optional stable adapter stores aggregates, summaries and histograms in one host-owned stable B-tree map using ic-memory's collection types. Namespaced records retain units, reset epochs, wide totals and overflow flags. Adding ic-metrics-stable is the opt-in; existing arithmetic callers remain volatile. The adapter requires Rust 1.88 on host and Wasm. Hosts own allocation/bootstrap, synchronization, write cadence and resets; the arithmetic core remains dependency-free and no_std on Rust 1.85.

The application guide is included in packaged crate documentation, with a compiled example of fixed named histograms, separate units and event counts. It covers sample admission, async attribution, resets, reporting and cost qualification; applications own those policies.

For a distribution, construct MeasurementHistogram::new([10, 100]). Its disjoint buckets cover 0..=10 and 11..=100; overflow() counts larger observations. bucket_counts() returns those two counts, and aggregate() includes every observation. Duplicate or descending bounds return a typed HistogramBoundsError identifying the invalid bound. Construction and recording support constant evaluation. Storage is fixed, recording searches at most the configured number of bounds, and every bucket saturates independently. Buckets describe ranges; they do not provide exact percentiles. Units, thresholds and export encoding remain consumer-owned. merge(&other) requires matching bounds and returns a typed mismatch without changing the destination. Bucket and overflow saturation flags distinguish exact caps from lost counts.

cumulative_count(index) provides exact counts through a configured bound and quantile_bucket(numerator, denominator) for nearest-rank ranges, such as p95 with (95, 100). Empty quantiles are None; required count saturation returns HistogramQueryError. Overflow has no configured ceiling. checked_scaled_ratio(numerator, denominator, scale) floors an exact scaled integer ratio from a u128 numerator, without floating point or overflowing intermediate multiplication. Callers establish exact inputs and handle empty or saturated diagnostics before using it. Consumer adoption remains separately qualified.

Known callers are IcyDB, Canic, ic-timers, ic-backup, IC Blob Storage's restoration test probe and Toko Miner's production game-shard action metrics. Their attribution, registries, callback roles, replication, persistence and reporting remain local. ic-backup's host integration measures local durations in nanoseconds and prepared chunk sizes in bytes, with a four-bound byte histogram in the inspected implementation. Its diagnostics establish no IC instruction or cycle cost. The crate has no global consumer registry, stable-memory allocation or endpoints. Consumers own their summary and histogram instances, labels and persistence policy. The extraction contract records scope and evidence.

Platform reads belong in consumer adapters: IcyDB and Canic use ic_cdk::api::call_context_instruction_counter(); IC Timers uses its existing ic0::performance_counter(1) binding. Both read counter 1. Consumers own target gating, native handling, call-context identity and attribution. IcyDB's inclusive overlapping spans and Canic's exclusive endpoint accounting remain different consumer contracts.

Measuring cost

Measure raw canister Wasm bytes, IC instructions and actual cycle charges separately. Shared arithmetic allocates nothing; registry lookups, labels, encoding and stable writes are separate consumer costs. Build profiles belong to each application, and native timing is not IC measurement evidence.

The tracked-recording comparison, quantile comparison and durable-store experiment retain reproducible sources, locked graphs, Wasms and raw IC readings. Their bounded workloads do not establish whole-application or mainnet savings. Original reader, histogram and consumer-lookup evidence remains unchanged under docs/evidence; the cleanup receipt retains the previous guide text. Compute distribution queries during reporting when possible.

Development

Common contributor commands use the reviewed Shared Tooling Make interface. Product builds, compiler floors, validation, publication and release metadata remain owned here.

Use rustup with the pinned Rust 1.99.0 toolchain, Git, GNU Make and Bash 3.2 or newer. Install wasm32-unknown-unknown for Wasm checks. The arithmetic library's minimum compiler is Rust 1.85.0.

The private Wasm evidence inspector requires Rust 1.88.0. It uses IC Host libraries to report raw/encoded bytes and structural counts for explicitly supplied Wasms with caller-selected bounds, without executing a canister. Bare Cargo commands select the arithmetic package.

Prepare and check the reviewed tools, select their paths, then install the repository's formatting hook:

make install-tools
make tools-check
export PATH="$PWD/.tools/host/bin:$PWD/.tools/ic/bin:$PWD/.tools/rust/common/bin:$PATH"
make install-hooks

Local setup owns bootstrap prerequisites, tool inventory, shared storage, installer failure handling and PATH selection. Git hook rules own staged formatting and hook installation. The reviewed shared tooling is vendored; normal commands need no Shared Tooling sibling checkout.

make fmt sorts workspace manifests and formats Rust; make fmt-check checks without editing. make help lists focused commands. Select named tests during implementation, and use make check, make check-wasm, make clippy and make wasm-inspect-check for the affected package. Completed code delivery uses make ci, make msrv, make stable-msrv and make wasm-inspect-msrv; inspection and documentation changes use narrower checks. The host support and qualification guide owns local gates and native acceptance. Agent rules own command authority.

Common IC executable setup and PocketIC ownership are documented in IC tools. Metrics has no active server caller or Testkit dependency. Repository maintenance procedures live in the task catalog.

Releases and adoption

Declare the current arithmetic contract in the consumer's root catalog:

[workspace.dependencies]
ic-metrics = "0.8"

Members inherit with ic-metrics = { workspace = true }. Published consumers resolve the registry package; consumers exposing shared types coordinate the same released dependency identity. The changelog owns release history and required hard-cut changes. The extraction contract and consumer qualification issue own integration scope and acceptance; the current handoff distinguishes local preparation, publication and hosted qualification.

Contributions follow the shared contribution rules. Maintainer releases use make release-patch, make release-minor or make release-major under the shared release contract. This repository supports RELEASE_DELIVERY=direct only, defaulting to remote origin and branch main. Release execution requires cargo-edit (cargo set-version). The shared contract owns preparation, validation, commit/tag delivery and interrupted-release recovery.

Publication is separate from release delivery:

make publish-check
make publish

publish-check is Cargo's dry run; publish uploads only the existing ic-metrics and ic-metrics-stable package versions to crates.io in dependency order, using normal Cargo credentials and locked dependencies. Both retain Cargo's package verification and dirty-tree rejection. Publication does not bump versions, create Git commits/tags or clean artifacts. See Cargo's publication documentation.